🤖 machine learning

The Kinetics of Training: A Driven-Nucleation Rate Law for Emergence, Plasticity Loss, and Circuit Control in Language Models

This paper proposes a driven-nucleation rate law explaining that language model capabilities emerge abruptly via a rare, all-or-nothing alignment of circuit components rather than gradual improvement, a mechanism that allows for precise prediction of capability ignition, diagnosis of learning failures, and targeted circuit repair through re-initialization.

Original authors: Lei Dong

Published 2026-07-31

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Lei Dong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a pot of water on a stove. You know the rules of thermodynamics: if you heat it enough, it will eventually boil. But knowing that it will boil doesn't tell you when the first bubble will break the surface, or how long you have to wait before the water stops being able to boil at all because the pot has become too full of something else. For a long time, scientists studying how AI models learn have been like people staring at that pot, describing the final state of the water (the "phase diagram") but missing the messy, ticking clock of how the bubbles actually form. They knew the water was hot, but they didn't have a stopwatch for the moment the AI suddenly "gets" a new skill. This paper steps into that gap, treating the training of an AI not as a smooth, steady climb, but as a frantic, high-stakes game of chance where a complex circuit must snap into place all at once.

The authors of this paper, Lei Dong, propose a new way to look at how language models learn new abilities. They argue that learning a new skill isn't like stacking bricks one by one, where you get credit for every brick you place. Instead, it's more like trying to solve a lock where you need to find the exact right combination of keys simultaneously. If you find all but one key, the lock doesn't open; you get zero credit. This "all-or-nothing" moment is the bottleneck. The paper introduces a mathematical formula, a "rate law," that predicts exactly when this lock will click open, how long it takes, and why sometimes, even if you keep training, the lock never opens at all. They call this the "Kinetics of Training," borrowing ideas from materials science (like how metal hardens) to explain how AI circuits are built, broken, and repaired.

The "All-or-Nothing" Lock

Imagine you are trying to teach a robot to recognize a specific pattern, like a secret handshake that requires five people to raise their hands at the exact same time. In the old way of thinking, you might expect the robot to get better as it learns to raise one hand, then two, then three. But this paper shows that for certain complex skills, the robot doesn't get any better until all five hands go up at once. If four hands go up and one stays down, the robot sees nothing. It's like trying to start a car with a key that has five cuts; if you get four cuts right but the fifth is wrong, the car won't start. You don't get a "half-start."

The authors call this the "no-partial-credit" rule. They measured this in their experiments and found that if a robot is missing three parts of a five-part circuit, it waits just as long to learn the skill as a robot missing three parts of a three-part circuit. The size of the circuit doesn't matter as much as the number of missing pieces. The "wait time" depends on how many pieces are missing, not how big the whole puzzle is. This changes everything: it means the hardest part of learning isn't the slow accumulation of knowledge, but the sudden, rare moment when all the missing pieces finally align by chance.

The Clock and the Floor

Because this alignment is a rare event, the authors built a "clock" to predict when it will happen. Think of it like a weather forecast for a lightning strike. You can't predict the exact second a bolt will hit, but if you know the humidity and the temperature, you can calculate the probability. The paper shows that by looking at a "precursor" signal—a tiny, early hint that the robot is starting to get the right pieces in place—you can predict exactly when the full skill will appear, often days before the robot actually starts performing the task well. In their tests, this clock predicted the arrival of new skills in six different AI models with a median error of just 5%.

However, there is a catch. The paper discovered a "floor" or a threshold. If the robot isn't exposed to the right kind of practice data often enough, it will never learn the skill, no matter how long you train it. It's like trying to grow a rare flower in a garden where the soil is too poor; you can water it for years, but if the soil lacks a specific nutrient, the flower will never bloom. The authors found that if the concentration of the right data drops below a certain point (around 0.26 in their specific tests), the learning rate drops to effectively zero. The robot doesn't just learn slowly; it stops learning entirely.

The Silent Death of Learning (Plasticity Loss)

One of the most striking findings is about what happens when a robot trains for too long on the wrong things. Imagine a student who spends years studying history but is suddenly asked to solve math problems. They might still be able to do the math, but the paper suggests that the ability to learn new math is slowly dying. The authors call this "loss of plasticity."

They found that as the robot trains, its internal "switches" (called attention heads) get "pinned" or stuck in specific patterns. Once these switches are stuck, the robot can no longer rearrange them to learn a new, complex circuit. It's like a room where all the furniture has been glued to the floor; you can't rearrange the room to fit a new party layout. The paper proves that this isn't because the robot is "full" of data, but because the specific parts needed to build the new circuit have been occupied and glued down by old habits.

Crucially, the paper shows that standard monitoring tools (like checking the robot's error rate) are blind to this. The robot's error rate might look like it's improving smoothly, while inside, its ability to learn new things is quietly dying. The authors even found a "deadline": if you wait too long to try to teach a new skill, it becomes impossible to learn, no matter how much you try.

The Cure: Resetting the Switches

But there is good news. The authors didn't just find the problem; they found the cure. They discovered that you don't need to retrain the whole robot to fix this. You only need to "reset" the specific switches that are stuck. In their experiments, they found that if they took an "aged" robot that had lost the ability to learn a new skill and simply reset the "query" and "key" parts of its attention mechanism (the parts that decide what to pay attention to), the robot instantly regained its ability to learn. It was like unlocking the glued furniture and letting the room be rearranged again.

Interestingly, resetting the other parts of the robot (the "value" parts) did nothing. This tells us exactly where the "glue" is: it's in the decision-making part of the circuit, not the memory part. This is a precise, targeted fix, not a brute-force reboot.

The Control Panel: Heating and Cooling

Finally, the paper treats the training process like a metallurgist treating metal. In metalworking, you can heat metal to make it soft (annealing), cool it quickly to make it hard (quenching), or hold it at a specific temperature to keep it in a certain state. The authors show that by injecting a specific kind of "noise" (randomness) into the training process, they can do the same with AI circuits.

They found that if you add just the right amount of noise, you can speed up the learning of difficult skills (the "nose" of the curve). If you add too much, you break the circuits. If you stop the noise at the right time, you can "pin" a circuit in place so it doesn't dissolve. They even showed that you can selectively "melt" (destroy) a complex circuit while leaving its simpler parts intact, like melting a chocolate sculpture but keeping the mold. This turns training from a black box into a controllable process where you can schedule when circuits form, when they break, and how to fix them.

What This Means

This paper doesn't just say "AI learns." It gives us a stopwatch, a thermometer, and a repair manual. It tells us that learning is a race against time and a battle against "glued" parts. It shows us that we can predict when a skill will appear, why it sometimes never appears, and exactly how to fix it when it breaks. The authors are careful to say that while they have proven these rules in controlled experiments and measured them in real models, the full mathematical proof for every type of AI is still being worked out. But for the models they tested, the rules are clear: learning is a rare, all-or-nothing event, and if you want to control it, you need to understand the clock, the floor, and the glue.

Technical Summary: The Kinetics of Training

Problem Statement

Current theoretical frameworks for language model training largely describe learning in equilibrium terms (e.g., Singular Learning Theory, phase transitions of free energy). While these models explain which phases exist, they fail to answer critical practitioner questions regarding rates: when will a capability emerge, how fast will it grow, and why does a network lose the ability to learn specific capabilities after prolonged training on irrelevant data (loss of plasticity)? The paper argues that training lacks a "time–temperature–transformation" (TTT) diagram—a kinetic framework analogous to metallurgy—that governs the formation and destruction of computational circuits.

Methodology

The paper constructs a rate equation for the nucleation of computational circuits, treating capability formation as a rare-event stochastic process driven by Stochastic Gradient Descent (SGD).

The Rate Equation:
The core object of study is the nucleation rate $J$ :
$J = N \times \nu \times \sigma(c) \times e^{-\beta K} - D$
Where:

$N$ : Number of available sites (recruitable attention heads).
$\nu$ : Attempt frequency.
$\sigma(c)$ : Supersaturation (concentration of capability-relevant statistics in the data stream).
$e^{-\beta K}$ : The barrier term, exponential in $K$ (the number of jointly necessary parts).
$D$ : Destruction term (drift or noise-induced damage).

Experimental Apparatus:
The authors utilize three primary experimental classes:

Controlled Toys: 2-layer GPT-NeoX transformers trained on synthetic tasks (e.g., $K$ -wise conjunctions, query-parity) where the "no-partial-credit" property is mathematically guaranteed. This allows for the isolation of the joint-alignment barrier.
Public Checkpoint Observation: Mechanism-level probes on Pythia and OLMo models (up to 1.4B parameters) to measure induction scores and precursor signals.
Controlled Continued Training: Aging substrates on plain data, then switching to a target data mix to measure emergence, or introducing "injected temperature" (isotropic weight noise) to test control protocols.

Key Assumptions:

Assumption 1: SGD behaves as a diffusion process with sufficient mixing on the formation timescale (the single unproven premise for the physical reading).
Assumption 2: Formation requires simultaneous joint alignment of all $K$ parts; partial progress on subsets yields no credit (no partial credit).
Assumption 3: Sub-alignment events are approximately independent (mean-field idealization).

Key Contributions

1. The Barrier Law and the "No-Partial-Credit" Principle

The paper proves that for tasks where no strict subset of parts carries signal (e.g., sparse parity), the formation barrier is additive in $K$ .

Finding: A circuit missing 3 parts waits as long as a smaller circuit missing 3 parts, regardless of the total circuit size.
Evidence: In a "missing-parts grid" experiment, formation times for circuits with $K-s=3$ missing parts were statistically identical (ratio 1.19–1.37) across total sizes $K=3, 4, 5$ . This confirms that the rate-limiting step is the joint alignment of the missing components, not the total circuit size.
Implication: The barrier grows exponentially with $K$ ( $e^{-\beta K}$ ), making high-order conjunctions significantly harder to learn than sequential assembly.

2. The Clock Law (Prospective Prediction)

The authors derive a "clock" that predicts the arrival of a capability based on its precursor.

Mechanism: The time to form a $K$ -part circuit is proportional to the time to form its $(K-1)$ -part precursor, scaled by a constant factor $e^{\beta}$ .
Result: Using a frozen constant $\beta \approx 0.23$ (measured on Pythia), the model predicted the emergence of induction capabilities in 6/6 held-out models (including OLMo-1B) with a median error of 5%.
Significance: This allows for the dating of capability arrival before the capability itself becomes measurable (when the performance curve is still flat at baseline).

3. The Floor (Supersaturation Threshold)

The paper establishes that capability formation is subject to a critical data concentration threshold, distinct from total token volume.

Finding: Below a critical concentration $c_0$ of capability-relevant statistics in the data mix, the formation rate is suppressed to effectively zero, regardless of training budget.
Evidence: In a concentration sweep, models trained with $c \le 0.26$ (70M scale) never ignited, while those with $c \ge 0.35$ always did. Increasing batch size could not compensate for low concentration.
Definition: Supersaturation is defined as the target's share of the gradient against a persistent competitor, not just token share.

4. Loss of Plasticity and the "Site" Term

The paper provides a quantitative law for why models lose the ability to learn new capabilities after "aging" on irrelevant data.

Mechanism: Training consumes the "site" term ( $N$ )—the pool of uncommitted attention heads. As heads become "pinned" to background tasks, $N$ decreases.
Horizon: There exists a "plasticity horizon" $t_c$ where $N$ drops below the minimum required to overcome destruction ( $N < N_{min}$ ). Beyond this point, increasing data drive cannot rescue learning.
Repair: The paper demonstrates that re-initializing only the Query-Key (QK) slices of attention heads (restoring the "score" channel) recovers learnability (6/6 success), whereas resetting Value (OV) slices or MLPs does not. This isolates the mechanism of plasticity loss to attention allocation patterns.
Blindness: Validation loss continues to decrease smoothly during this aging process, rendering standard monitoring blind to the loss of plasticity.

5. Process Control via Injected Temperature

Since native SGD lacks a valid "temperature" (failing the fluctuation-dissipation test), the authors propose injecting isotropic weight noise to complete the kinetic equation.

The Nose: A "TTT nose" exists where a specific temperature minimizes formation time by assisting barrier crossing without triggering destruction.
The Melt Line: Destruction follows a "weakest-link" law; complex circuits melt at lower noise thresholds than their constituent parts.
Pinning and Release: Circuits can be pinned in a sub-critical state and released via "quench plus surgery" (rescaling attention weights), converting stored heat into a head start for learning.

Results Summary

Emergence: Validated a prospective clock predicting capability arrival with 5% median error across model families and scales.
Plasticity: Quantified the "occupation tax" (depletion of sites) and demonstrated a causal repair mechanism (QK reset) that restores learnability in aged models.
Thresholds: Proved the existence of an intensive data-mixing floor below which capabilities cannot form, regardless of compute budget.
Control: Demonstrated that injecting temperature allows for the acceleration of high-barrier circuits (annealing), selective dissolution of circuits (melting), and the pinning of sub-critical structures.

Significance and Claims

The paper claims to provide the first "kinetics" of training, moving beyond static phase diagrams to a dynamic, rate-based framework.

Unification: It unifies emergence, plasticity loss, and circuit control under a single rate equation derived from the stochastic dynamics of SGD.
Predictive Power: It offers a method to predict when capabilities will emerge and when they will become unlearnable, addressing gaps in current interpretability and safety monitoring.
Causal Mechanism: It moves beyond correlation to identify the specific causal mechanism of plasticity loss (attention subspace commitment) and a targeted repair strategy.
Scope Limitations: The authors explicitly note that the "mixing" assumption (Assumption 1b) remains unproven for transformer SGD, meaning the thermodynamic interpretations (temperature, melting) are conditional on this premise. The results are currently validated on models up to 1.4B parameters and specific capability classes (conjunctions, induction), not yet on full-scale production benchmarks.

The paper concludes that while the mathematical structure relies on specific assumptions, the empirical constants and kinetic laws hold across different model families, suggesting that the "rate-determining step" of joint alignment is a fundamental property of learning in these architectures.